DEV Community

AI-KSK
AI-KSK

Posted on

LTX 2.3 / 3.5 Production Suite in ComfyUI: 10 Workflows for Multi-Image Reference, Retalk Lip-Sync, Storyboards, Expansion and Watermark Removal

#comfyui #ai #tutorial

This is a production-oriented LTX suite rather than a single effect: ten workflow files covering character consistency, talking-head replacement, storyboard direction, frame expansion, subtitle/watermark cleanup and first-frame-similar video.

Evidence note: ten project files were announced, eight have published RunningHub pages (verified); the remaining two are described but not linked in the resource list. No workflow JSON was parsed for parameters.

The eight verified routes

Multi-image lock / MSR three-stage reference (https://www.runninghub.ai/post/2068718418597081089/?inviteCode=rh-v1111) — locks a character across a sequence with multi-image reference and a three-stage render. Use it when wardrobe/face drift is the failure mode.

RetalkPro — change the voice, not the person (https://www.runninghub.ai/post/2068718437811179522/?inviteCode=rh-v1111) — re-dubs a video while keeping the performance. The practical use is localisation: same actor, new language.

Role separation / storyboard director (https://www.runninghub.ai/post/2068718478793728001/?inviteCode=rh-v1111) — separates character, action and scene control, the structure you need for multi-shot pieces.

Six-grid storyboard director (https://www.runninghub.ai/post/2068718977450340353/?inviteCode=rh-v1111) — lays shots out as a grid so continuity is reviewable before rendering.

Avatar opening — single-person digital human (https://www.runninghub.ai/post/2068714653911445505/?inviteCode=rh-v1111) — image + audio driven talking/singing head.

Intelligent frame expansion (https://www.runninghub.ai/post/2067822318637895681/?inviteCode=rh-v1111) — changes aspect ratio without cropping the subject.

Subtitle / watermark removal (https://www.runninghub.ai/post/2067822291773378561/?inviteCode=rh-v1111) — cleanup pass for clips that carry burned-in text.

First-frame high-similarity video (https://www.runninghub.ai/post/2067652016737968129/?inviteCode=rh-v1111) — keeps the opening frame close to a given still, which matters when a video must match a poster.

Why this suite is different

Each graph solves one production problem, and they share the LTX 2.3/3.5 three-stage render and LoRA-weight control conventions. That means you can hand off between graphs — e.g. expand the frame after a storyboard pass — without relearning the pipeline.

Suggested order for a real piece

  1. Storyboard with the six-grid director.
  2. Lock the character with MSR.
  3. Render shots; keep the first frame matched.
  4. Expand to the delivery aspect ratio.
  5. Re-dub/Retalk for localisation.
  6. Clean subtitles/watermarks last.

Limits

  • Two of the ten announced workflows have no published link in the resource list.
  • No JSON/parameters were parsed; step counts are not quoted.
  • LTX behaviour depends on model build and node versions.

Episode resources

Top comments (0)